Iryna Schlackow and Borislava Mihaylova
on behalf of the SHARP Collaborative Group
useR! 2019
Toulouse, July 11, 2019
Facilitating external use with user-friendly
interfaces: a health policy model case study
Motivation:
what is a health policy model?
A health policy model is a tool to inform policy decisions by
projecting people‘s life courses. Predictions include
• disease events
• life expectancy
• quality of life
• healthcare costs
• effects of treatments
- positive (disease risk reduction) and negative (adverse effects)
Projections made over long time periods (eg lifetime)
Motivation:
why are health policy models needed?
Healthcare budgets are limited and not all treatments can be
recomended even if effective
• Models show whether treatments are good value for money
• Health policy models are increasingly used by policy makers
and clinicians
• In UK, cost-effectiveness analyses are required by NICE
- Good-value-for-money: £20-30K per extra quality-adjusted life-year
(QALY)
• Flexible models can help answer many policy questions
• Aim for transparency, reliability, reproducibility and usability
Motivation:
how to facilitate usability?
Transparency Reliability Usability
Release
the code
😀 😀
☹️
useRs only
code mis-use
Publish equations
and
methods
🙂 🙂
☹️
analysts only
Provide user
-friendly
interface
☹️
black box
☹️
😀
NB: user vs useR
Publish equations and
methods and provide
user
-friendly interface
🙂 🙂 😀
SHARP CKD-CVD model:
Shiny interface
Case study: SHARP CKD-CVD model
Background
• Chronic kidney disease (CKD) increases cardiovascular (CV) risk
• Want to project long-term outcomes in CKD
– cardiovascular events, CKD progression, life expectancy, quality of life,
healthcare costs;
– enable implementation of treatments to reduce cardiovascular risk
∙ assess long-term effects and cost-effectiveness.
• Patient-level data from a trial
– baseline characteristics, within-trial events
• Risk equations derived from the data
• Combined into a Markov model to do lifelong projections
– validated internally and externally
SHARP CKD-CVD model:
need for a user-friendly interface
• The model to be useful for NICE, other analysts, clinicians...
• User-friendly interface accessible from anywhere
• No need for knowledge / installation of R
• Adaptation to other scenarios/countries
- national mortality rates
- national healthcare costs
• Customising parameters in the current setting
- treatment to be assessed
- population characteristics
- duration of treatment / time horizon
- discount rate
SHARP CKD-CVD model:
Shiny interface
http://one-elevenbooks.com/shiny-or-the-truth/
http://dismod.ndph.ox.ac.uk/kidneymodel/app/
• Application accessed via a link
• The user only sees the front end
• All programs/data stored externally
• The front end can be modified using CSS
themes, htmlwidgets, and JavaScript actions
- fancy fonts, links, email addresses etc
- error checking on data entry
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
SHARP CKD-CVD model:
Shiny interface
User-friendly interface:
help with debugging and transparency
User-friendly interface:
help with debugging and transparency
• Face validity debugging
- Easier to do on a user-friendly interface (even for the developers!)
• Feedback from external users
• Running several models against a reference simulation
- Mount Hood diabetes challenge: models predicting long-term outcomes in
diabetes patients
• everyone gets the same tasks (eg change in life expectancy after statin initiation)
• core assumptions same for everyone
• additional assumptions must be documented in a pre-defined template
• the results are presented, compared and (usually) published
• user-friendly interface enables replication
SHARP CKD-CVD model:
conclusions
• SHARP CKD-CVD model is a novel resource for evaluating health
outcomes and cost-effectiveness of interventions in CKD
• User-friendly web-based freely available interface aids model use
• Together with the published equations / methods helps ensure
reliability of the underlying code and methods transparency
• The user can enter with their own parameter values and perform
calculations in different settings
• User’s perspective taken into account:
- simple menus, straightforward navigation, pretty looks
- detailed user-guide
- example input/output files, file descriptions and default values
- error checking at data entry could (partially) prevent inappropriate use
- which parameters should be modifiable?
SHARP CKD-CVD model:
challenges and discussion points
• Day-to-day support
- Replying to queries, fixing bugs
- R/package updates may break everything!
- Not updating is not an option (according to our IT team)
• Is R the best option for such an interface?
- Might Python be faster and/or have better visualisation capabilities?
- C++?
• Do the benefits of releasing the code outweigh the risks?
Acknowledgements
Seamus Kent, Richard Haynes, Jonathan Emberson, Will Herrington, Colin
Baigent, Alastair Gray, Jingky Lozano-Kuehne, Martin Craig, Martin
Landray, Kirsty Reith
SHARP participants, study staff and collaborators!
The SHARP study was funded by Merck/Schering- Plough Pharmaceuticals
(North Wales, PA, USA), with additional support from the Australian
National Health Medical Research Council, the British Heart Foundation,
and the UK Medical Research Council
